Counting Activities Using Weakly Labeled Raw Acceleration Data: A Variable-Length Sequence Approach with Deep...

Area of Science:

• Human Activity Recognition
• Machine Learning
• Wearable Sensor Technology

Background:

• Counting hand-performed activities is challenging due to variable activity durations.
• Traditional fixed window sizes lead to inaccurate activity representation.
• Weakly labeled data simplifies annotation but poses challenges for machine learning.

Purpose of the Study:

• To develop a novel deep learning approach for counting hand-performed activities using IMUs.
• To address the limitation of fixed window sizes by segmenting data into variable-length sequences.
• To leverage weakly labeled data for simplified annotation and reduced data preparation time.

Main Methods:

• Utilized ragged tensors for segmenting time series data into variable-length sequences.
• Implemented a Long Short-Term Memory (LSTM)-based deep learning architecture.
• Employed weakly labeled data, providing only partial information about performed activities.

Main Results:

• Achieved a repetition error of ±1, even in challenging cases, on the Skoda HAR dataset.
• Demonstrated the effectiveness of variable-size IMU acceleration data processing.
• Showcased a computationally efficient approach for activity counting.

Conclusions:

• The proposed method effectively counts hand-performed activities using variable-length sequences and weak labels.
• This approach offers a computationally efficient solution for activity counting with IMUs.
• Findings have broad applications in healthcare, sports, HCI, robotics, and manufacturing.